This should be an obvious corollary. Transformers scale exponentially with context tokens, and it was well-known current tokenizers perform worse on non-English text, splitting it into more tokens.